lecture notes #4

advertisement
Advanced Algorithms 2012A
Lecture 4 – flow/cut gaps∗
Robert Krauthgamer
1
Maximum Flow and Minimum Cut (Single Commodity)
Let G = (V, E) be a directed graph with a source s ∈ V , a sink t ∈ V , and edge capacities
ce ≥ 0. Recall that maximum st-flow is equal to minimum st-cut. We first review material from
the previous class, in which Uri gave an alternative proof using LP duality and total unimodularity.
Maximum flow can be written as the following LP with variables (xuv : uv ∈ E ∗ ). We define
E ∗ = E ∪ {ts} which adds an edge from t to s.
maximize xts∑
subject to
v:
∑
xvu −
vu∈E ∗
xuv ≤ cuv
xuv ≥ 0
v:
xuv = 0 ∀u ∈ V
uv∈E ∗
(1)
∀uv ∈ E
∀uv ∈ E ∗
The dual LP has variables (yu : u ∈ V ) and (yuv : uv ∈ E) and is given by:
minimize
∑
cuv yuv
uv∈E
subject to yu − yv − yuv ≥ 0 ∀uv ∈ E
yt − ys ≥ 1
yuv ≥ 0
∀uv ∈ E
(2)
This LP is a relaxation of the min-cut problem. We can call it “fractional cut”. Using also duality,
we thus have
max-flow = fractional-cut ≤ min-cut
Exer: prove the inequality max-flow ≤ min-cut directly by definition.
∗
These notes summarize the material covered in class, usually skipping proofs, details, examples and so forth, and
possibly adding some remarks, or pointers. The exercises are for self-practice and need not be handed in. In the
interest of brevity, most references and credits were omitted.
1
1.1
Using Total Unimodularity
We saw last week how to use TUM to prove that
max-flow = fractional-cut = min-cut.
1.2
Alternative LP formulations
Let P be the collection of simple paths from s to t. Here is a different but equivalent formulation
of max-flow using flow paths. It has a variable xp for every path p.
maximize
subject to
∑
p∈P fP
∑
fp ≤ ce ∀e ∈ E
(3)
p∈P :e∈P
fp ≥ 0
∀p ∈ P
Lemma: This LP is equivalent to LP (1) in the sense that they always have the same value.
Exer: Prove this lemma, by basically converting every feasible solution to one LP into a solution
for the other LP.
The dual to LP (3) given below has variables (ye : e ∈ E).
minimize
subject to
∑
e∈E
∑
ce ye
ye ≥ 1 ∀p ∈ P
e∈P
ye ≥ 0
(4)
∀e ∈ E
This LP is an alternative relaxation of minimum cut.
Lemma: For every feasible solution ⃗y to LP (4) there is A ⊂ V , i.e. a cut (A, Ā), such that
∑
capacity(A, Ā) ≤
ce ye .
e∈E
Proof: seen in class, by interpreting ye as edge lengths, computing shortest-paths distances from s
and cutting at a random radius around s.
This gives yet another proof that
max-flow = fractional-cut = min-cut
KEY point: The dual of a flow problem involves distances, and distances are just “fractional”
versions of cuts.
2
Exer: Turn the above proof into an algorithm that converts a solution to LP (4) into an st-cut
with same value. What is its runtime?
Exer: Use a similar argument for the first relaxation of min-cut, i.e. prove that every solution to
LP (2) can be converted into an st-cut with no larger value.
Exer: Does this proof work for undirected graphs?
2
Multicommodity flow and multicut
Let G be an undirected graph with edge-capacities, and for each i ∈ [k] let {si , ti } be a pair of
vertices called source-sink or demand pairs. The maximum multicommodity flow asks to ship the
maximum amount of total flow, while respecting the edge capacities (the total flow on every edge
is at most the edge’s capacity).
The problem can be written as the LP below. We let Pi be the set of all si − ti paths.
maximize
subject to
∑ ∑
fi
∑i p∈P
∑i p
fpi ≤ ce ∀e ∈ E
i∈[k] p∈Pi :e∈p
fpi ≥ 0
(5)
∀i ∈ [k], ∀p ∈ Pi
In the multicut problem, the input is as above, and the goal is to find a minimum-capacity set of
edges E ′ ⊂ E such that removing E ′ from the graph disconnects every pair si − ti .
Exer: Give an example where G \ E ′ might have several connected components (for an optimal E ′ ).
Exer: show that in every network
maximum multicommodity-flow ≤ minimum multicut,
and give an example where the inequality is strict.
The dual of LP (5), given below, has variables (ye : e ∈ E).
minimize
subject to
∑
e∈E
∑
ce ye
ye ≥ 1 ∀i ∈ [k], ∀p ∈ Pi
(6)
e∈p
ye ≥ 0
∀e ∈ E
Observe that this LP is a relaxation of the multicut problem.
The flow/cut gap Theorem:
minimum multicut ≤ O(log k) · maximum multicommodity-flow.
3
The proof idea is to again interpret the ye variables as edge lengths, compute the shortest-path
distance d(u, v) between every two vertices, and “cut” around each ti (or si ). It relies on the
following randomized for partitioning the vertices. For a partition P of V , we shall call each S ∈ P
a cluster.
Algorithm CKR/FRT
Input: metric d on V , a set T ⊂ V of k vertices (called terminals), and parameter R > 0.
Output: a partition P of V
1. Choose a uniformly random permutation π of the terminals T
2. Choose β ∈ (1, 2) uniformly at random.
3. Iteratively for t = π(1), . . . , π(k), create a new cluster St containing the yet unclustered vertices
u with d(t, u) ≤ βR.
4. Let P consist of these St plus all the remaining vertices as an additional cluster V0 .
Notice that t need not belong to “its own” cluster St .
Lemma: Every cluster St ∈ P other than V0 has diameter < 4R.
Proof: d(u, v) ≤ d(u, t) + d(t, v) ≤ 2βR.
Lemma: For every u, v ∈ V ,
Pr[u, v are separated in P ] ≤ O(log k) · d(u, v)/R.
Proof: seen in class, by analyzing the events of the form: a center t “captures” u but not v.
Remark (important for later): A more careful examination of the proof shows that
Pr[u, v are separated in P ] ≤
|BT (u,3R)|
d(u,v)
R O(log |BT (u,R/2)| ).
The theorem can now be proved easily using the CKR algorithm.
Remark: It’s not hard to verify that this gives a polynomial-time O(log k) approximation for the
multicut problem, which is NP-hard even with k = 3.
Exer: Show a constant factor approximation for the multiway cut problem, which is a special case
of multicut
( ) where there is a set T of k vertices (called terminals), and the demand pairs {si , ti } are
just all k2 pairs in T .
4
Download